sg%ddlZddlmZddlZddlZddlZddlm Z m Z dZ dZ dZ dZedd d gZed d d dddZedd dddZedd dddZiej&j)Dcic]}|ec}ej&j)Dcic]}|ec}ZdZdZdZdZedk(reyycc}wcc}w)N) namedtuple)COMPUTE_METADATA_SCOPE_NAME TritonHookcg}|rl|D]e}|j}|D]N}|jddjj}|||fvs:|j|dzeg|S|dgdz}|S)N(r (inc))lowersplitstripappend)metrics raw_metricsretmetric raw_metricraw_metric_no_units I/var/www/html/venv/lib/python3.12/site-packages/triton/profiler/viewer.pymatch_available_metricsr s C F\\^F)  %/%5%5c%:1%=%C%C%E%K%K%M"j*<==JJzH45    J1~) Jctj|}|jd}tjj |}||j |fS)N)jsonloadpopht GraphFrame from_literalshow_metric_columns)filedatabase device_infogfs rget_raw_metricsr#sFyyH,,q/K  # #H -B r%%' 44rc ftjd|jdg}|D]}||D]}|||d}|||d}|||d}tjD]}|d|k(} || } d|| j vr"d } |d k(r9|d k(r d |d z z } n_|dk(r d|d z z } nQ|dk(rL|dz |zdz dzdz|d z z } n4|dk(r!|dk(r d|d z z } n!|dk(s|dk(rd|d z z } nt d||j| dfxx| d|jd | z z cc< |S)Nmin_timeindexcolumnsarchnum_sms clock_rateDeviceIdflopsrCUDA80g2C89gbB90rgx:AimBHIPgfx90ag//bBgfx941gfx942g?y"CzUnsupported device type: ) pd DataFramer(r flops_widthr) ValueErrorlocfillna) dfr!min_time_flops device_type device_indexr*r+r,widthidx device_frames max_flopss rget_min_time_flopsrH s\\#RXX |LN"l ' 4 lL{+L9&AD!+.|) r@r!min_time_bytesrBrCrErFrJrKpeak_bandwidths rget_min_time_bytesrPAs\\#RXX |LN"[ ' 4 [LZ.L0CsGM +K 8 FGZ [ #K0>{KI]->>DqHN   sJ /=3IN3Z Z /  [[ r FactorDictnamefactortimergMbP?gư>g& .>)time/sztime/msztime/usztime/nsr.geAr5)zflop/szgflop/sztflop/srM)zbyte/szgbyte/sztbyte/scg}g}ttjg|d}tjdz|jddjddz}|D]}|dk(rt |j |} t |j |} |j |tj|tjdz z} | dj| dt| z |j d <|jd |tvrt|} | j} | j}t| g|d}|j ||j |tj|zz ||z |j |d <|j|d B|tjvrtjdz|jddz}|j |tj|tj|z z|j |d <|j|d |j||r t||}||zS) Nr/rr)utilrUr&z util (inc)r) rtime_factor_dictrRr rP dataframerHrScombinemaxr derivable_metrics)r"r rr!derived_metricsoriginal_metricstime_metric_name time_unitrrNrAtime_secderiveable_metric metric_namemetric_factor_dictmatched_metric_namemetric_time_units rderive_metricsri[spO.0@0E0E/F TUVW!&&,/?/E/Ec/J1/M/S/STW/XYZ/[[I, V / kJN/ kJN||$459I9P9PQZ9[9I9P9PQY9Z:[\H)7 )C)K)KN[eLfhk)low)wBLL &  " "< 0 ( ( 1& 9 +00K!2!9!9 "9;-"UVW"X /1||= F) metric_column expand_namedepth render_header) openr#lenupdate_inclusive_columnsrifilterrprinttree) r filenameincludeexclude thresholdrpfr"rr! name_filterquerys rparser~s h b'6q'9$K;!#I%II# ##% Wk;G g18D -gK6;/0E5.BD!< =SABC YYuTY * 71:YK'89:E5.B bggGUZ_g`a#bbbs C CC(c t|d5}t|\}}}td|rE|D]@}|jddj j }td|B dddy#1swYyxYw)NrkzAvailable metrics:rrz- )rrr#rvr r r ) file_namer|_rrrs r show_metricsrs i +A.; "# ) 1 %/%5%5c%:1%=%C%C%E%K%K%M"-./0 1 s A#A99Bctjdtj}|jdddd|jdd tdd |jd d tdd |jddtdd |jddt dd |jddt dd |j\}}t|dk(sJd|d}|jr|jjdnd}|j}|j}|j}|j}|r |r td|j r t#|y|rt%||||||yy)Nz,Performance data viewer for proton profiles.) descriptionformatter_classz-lz--list store_truea`List available metrics. Metric names are case insensitive and ignore units. Derived metrics can be created when source metrics are available. - time/s, time/ms, time/us, time/ns: time - flop/s, gflop/s, tflop/s: flops / time - byte/s, gbyte/s, tbyte/s: bytes / time - util: max(sum(flops) / peak_flops_time, bytes / peak_bandwidth_time)) )actionhelpz-mz --metricszAt maximum two metrics can be specified, separated by comma. There are two modes: 1) Choose the output metric to display. It's case insensitive and ignore units. 2) Derive a new metric from existing metrics. )typedefaultrz-iz --includez?Include frames(kernels) that match the given regular expressionz-ez --excludez?Exclude frames(kernels) that match the given regular expressionz-tz --thresholdzrExclude frames(kernels) whose metrics are below the given threshold. This filter only applies on the first metric.z-dz--depthdz The depth of the tree to displayrzMust specify a file to readr,z'Cannot specify both include and exclude)argparseArgumentParserRawTextHelpFormatter add_argumentstrfloatintparse_known_argsrsr r ryrzr{rpr=listrr) argparserargs target_argsrr ryrzr{rps rmainrs''B 55I          N    N   |    / "224D+ { q ?"?? AI)-dll  %4GllGllGI JJE7BCC yyY  gy'7IuE r__main__)r collectionsrrpandasr:hatchetrtriton.profiler.hookrrrr#rHrPrQrZflops_factor_dictbytes_factor_dictrSkeysr^rirrr__name__)keys0rrs#" H 5B  vx&8 9 ftPTae&fgw1QU(VWw1QU(VW#**//13  3$**//13  3 .Fb*JFZ zFo33s ; C $ C